DISCOVER THE FUTURE OF AI AGENTSarrow_forward

AWorld

calendar_todayAdded Feb 25, 2026
categoryAgent & Tooling
codeOpen Source
PythonWorkflow AutomationDockerPyTorch大语言模型Multi-Agent SystemMultimodalAI AgentsReinforcement LearningAgent FrameworkCLIAgent & ToolingOtherAutomation, Workflow & RPAModel Training & Inference

An open-source framework for building, evaluating, and training general multi-agent systems. Features natural language agent creation, distributed reinforcement learning training pipeline, and complex environment interactions. Ranks top on authoritative benchmarks including GAIA, OSWorld, and VisualWebArena.

Overview#

AWorld is developed by the Inclusion AI team with the core philosophy "The Next Frontier for AI is Your Expertise", enabling domain experts to build their own fleet of autonomous agents.

Core Capabilities#

Agent Creation & Evolution

  • AWorld-CLI: Command-line tool supporting automatic agent code generation from natural language descriptions
  • Three Evolution Paths: Automated creation, manual code editing, AI-assisted optimization (built-in Optimizer Agent as pair-programming partner)
  • Verified Skills Library: Curated skill blueprints including PPT Agent (presentation generation) and DeepSearch Agent (multi-source research)

Multi-Agent Orchestration

  • Flexible collaboration topology
  • Parallel task runtime
  • Streaming response support
  • Human-in-the-loop mechanism
  • Comprehensive observability tracing

Environment Interaction

  • Rich toolset (web browsing, computer control)
  • Secure sandbox environment
  • MCP (Model Context Protocol) support

Reinforcement Learning Training

  • Distributed task execution with 14.6x speedup vs single-node
  • Large-scale agent-environment interaction experience collection
  • Complete Agentic AI training pipeline

Benchmark Performance#

BenchmarkScoreRank
GAIAPass@1: 67.89, Pass@3: 83.49Open-source #1
OSWorld58.0% success rate#1 (2025/09)
VisualWebArena36.5% success rate#1 (2025/09)
XbenchPass@1: 51, Pass@3: 61Excellent

Application Scenarios#

  • Scientific Research: IMO 2025 math problem solving (5/6 problems within 6 hours)
  • Web Automation: VisualWebArena benchmark leader
  • Computer Control: OSWorld #1
  • Deep Search: Excellent Xbench performance
  • General Assistant: GAIA open-source leaderboard top

Quick Start#

# Clone repository
git clone https://github.com/inclusionAI/AWorld && cd AWorld

# Create environment
conda create -n aworld_env python=3.11 -y && conda activate aworld_env

# Install
pip install -e . && cd aworld-cli && pip install -e .

# Launch CLI
aworld-cli

Configuration#

Create .env file in AWorld/aworld-cli directory:

LLM_MODEL_NAME="Claude-Sonnet-4 or above suggested"
LLM_PROVIDER="openai"
LLM_API_KEY="your_model_api_key"
LLM_BASE_URL="your_base_url"

Project Structure#

AWorld/
├── aworld-cli/     # CLI tool
├── aworld/         # Core framework
├── docs/           # Documentation
├── env/env_channel/# Environment channel
├── examples/       # Example code
├── train/          # Training modules
└── tests/          # Tests

Target Users#

  • Domain experts (scientists, financial analysts, engineers)
  • AI researchers
  • Developers requiring customized agent solutions

Related Projects

View All arrow_forward

STAY UPDATED

Get the latest AI tools and trends delivered straight to your inbox. No spam, just intelligence.

rocket_launch